1. 44
  1.  

  2. 10

    I’ve also been totally consumed by the same obsession. Lots of money, too: I want to be able to distribute VQGAN models on multiple gpus, which is far beyond my minimal pytorch knowhow. So I’m now looking to pay a contractor.

    I have this dream of making 2000x2000 pixel murals and printing them to canvas. AWS has EC2 configs with 96 gigs of gpu. I can’t stop thinking about this, and it’s disrupting my life.

    But it’s also exhilarating. I know it’s “just” an ai generator, but I’m still proud of the stuff I “make”. Here are some of my favorites:

    My daughter wants to be an artist. What should I tell her? Will this be the last generation of stylists, and we’ll just memorize the names of every great 20th century artist to produce things we like, forever?

    I worry about this too, but also am excited to see what artists do when they have these tools. And I think it’ll make artists turn more to things like sculpture and ceramics and other forms of art that are still out of the machine’s reach.

    EDIT: also, a friend and I have been making games based off this. “Guess the city this from this art” or “guess the media franchise”. It does really funny stuff to distinct media styles, like if you put in “homestar runner”

    1. 5

      Just my random observation but “your” pieces and the post’s all give the vague appearance of something running through endless loops of simulacra. Said another way, they all share similar brush strokes.

      I think we’re headed into the (while looking at a Pollock) — “humph, my AI could have painted that!” era

      1. 4

        There are a bunch of known public Colab notebooks but one is very popular. It’s fast but has this recognizable “brush stroke” indeed. Some GAN jockeys are tweaking the network itself though, and they easily get very different strokes at decent speeds. You don’t even need to know neural network math to tweak, just the will to dive in it. Break stuff, get a feel for what you like. If this is to become a staple artist’s tool it’ll have to be like that, more than just feeding queries.

      2. 3

        These are cool. The “Old gods” one especially… if that was hung in your house and you told me you’d purchased it from an artist I wouldn’t blink. When you make them, are you specifying and tweaking the style, and then generating a bunch, and then hand-picking the one you like?

        1. 3

          Starting out I was just plugging intrusive thoughts into colab to see what I’d get. If it didn’t produce something interesting (not many do) I’d try another prompt. Recently I spent a lot of time writing a “pipeliner” program so I can try the same prompt on many different configs at once. I got the MVP working on Monday, but I’m putting it aside a while so I can focus on scaling (it only works on one GPU, so can’t make anything bigger than 45k square pixels or so)

          1. 1

            Are you saying you’ve managed to get this to run locally? All the guides I’ve found are simply how to operate a Colab instance.

            1. 2

              I got it running locally, but I don’t have a GPU so upload it to an EC2 instance. I recently found that SageMaker makes this way easier and less burdensome, though.

        2. 1

          There are neural nets intended specifically for upscaling images. Pairing one of these with VQGAN image generation (which is pretty low res) might let you make larger scale art without a huge unaffordable GPU.

        3. 5

          The feeling I have is that AI may do for art and music what it has done for playing Chess and Go. We will perform for an audience of other humans, who are tickled to see a human try to do something that’s even a pale imitation of what a pile of linear algebra can do.

          Perhaps for muzak (and the endless stream of “chillout” compositions Spotify commissions to avoid paying songwriter royalties) as well as wallpaper for web pages and incidental illustration. But the same fears were expressed when photography essentially destroyed the portrait painting business in the 19th century. The output of a creative human still has value.

          1. 6

            I can see a big advantage here for stock photos. Currently, I have a choice of using an off-the-shelf image very cheaply and putting up with the fact that it’s only an approximate fit for what I want or paying a professional artist of some kind to produce something that’s exactly what I want. This gives a middle ground where I can get something that’s very close to what I want (these elements, this style, and so on). My guess is that it’s far more disruptive of the stock photo industry than it is to professional artists. There’s still a lot of value in having access to a model that’s trained on a particular artist’s works so that it can mimic their style better.

            1. 6

              Oh, definitely. This will absolutely destroy stock imagery as a business (not that it was in that great shape to begin with). Also, those CG interiors everyone now uses in catalogs can be “personalized” with unique posters/artworks, flower arrangements etc.

              Edit art qua art isn’t in trouble here, what’s going to be hit hard is “workman’s art” - stock photography, logo design, maybe even wedding photography, those insipid landscapes beloved of framing companies… and we’re not even getting into the coming plague of realistic-looking social media profiles.

              1. 2

                and we’re not even getting into the coming plague of realistic-looking social media profiles.

                If there is one profession that I would love to see automation completely and utterly destroy then ‘social media influencer’ is probably the one for which the phrase ‘and nothing of value was lost’ could be best applied to.

                1. 3

                  No argument there… but I’m thinking of bots controlled by a single entity but each one with a convincing avatar and grammatically correct bio, etc. In fact, I believe such bots are routinely used by the influencers to game ad spending.

          2. 2

            Absolutely phenomenal! Gurney’s Color and Light is one of my favorite books and these paintings (they are paintings, right?) captured some aspects of his style despite being formless blobs when examined carefully.

            1. 1

              While we’re on the topic… many years ago I played with a random raytracer program that generated raytraced images with flares etc. I enjoyed picking out the good outputs and saving as wallpaper, but now I have lost the binary and cannot find it online. It was for Linux/Unix, does anyone remember something similar?

              1. 1

                The feeling I have is that AI may do for art and music what it has done for playing Chess and Go.

                Indeed, ruin them.

                1. 5

                  I don’t think that’s fair. AIs are better than humans in chess, but it’s widely accepted that they’ve made human-level play much better, by making chess training and board analysis widely available even to people who can’t afford coaches.